HEALTH ECONOMICS
Sponsored by a Grant TÁMOP-4.1.2-08/2/A/KMR-2009-0041 Course Material Developed by Department of Economics,
Faculty of Social Sciences, Eötvös Loránd University Budapest (ELTE) Department of Economics, Eötvös Loránd University Budapest
Institute of Economics, Hungarian Academy of Sciences Balassi Kiadó, Budapest
2
Authors: Éva Orosz, Zoltán Kaló and Balázs Nagy Supervised by Éva Orosz
June 2011
Week 12
Modelling approaches in health economic evaluations
Authors: Zoltán Kaló and Balázs Nagy Supervised by Éva Orosz
Full Economic Evaluation:
assessment of cost-effectiveness
3
The role of modelling
WOSCOPS – Pravastatin in primary prevention WOSCOPS – clinical study
• 225 life years gain over 5 years among 10,000 patients with no history of CHD
• Cost per life year gained is: 100,000 GBP (not cost-effective) WOSCOPS - modelling
• Hypothesis: incremental life years still grow after 5 years (those who gained will not die shortly after their last visit, and avoided myocardial infarct also increases life expectancy)
• Cost per life year gained: 8–20,000 GBP (cost-effective)
Caro et al. BMJ. 1997. 315. 1577-82
Pharoah; Freemantle; Caro - BMJ 1998. 316. 1241-42.
Tools of economic evaluations
• Burden of Illness:
– Understand cost structure of disease in a given country – Define economic value hypotheses
– Demonstrate of importance of disease areas
• Economic evaluation alongside clinical trials
– Analyse economic consequences of efficacy and safety
• Naturalistic economic trials
– Analyse the cost-effectiveness under naturalistic conditions
• Economic modelling on the basis of clinical trial data
– Extrapolate to longer time horizon or specific populations/countries
4
Advantages Disadvantages
Piggy-back economic evaluation
• randomization (internal validity)
• low cost of economic data collection
• results available before reimbursement decisions
• selected patient population
• protocol driven costs
• limited time-frame
• poor monitoring health economic data compared to efficacy and safety data
• statistical power adjusted to efficacy end-points
• economic events after efficacy end-points
Naturalistic economic evaluation
• average patients in real life conditions (external validity)
• real costs (independent from trial protocol)
• simple implementation if patient ID available in payers’
or managed care database
• difficult data collection and monitoring
• limited options for randomization selection bias
• limited time-frame
• results after reimbursement decisions
Economic modelling
• generalizability, adjustment to local practice and population
• appropriate timing for major decision-points (e.g. pricing, reimbursement decisions)
• correct results only if modelling assumptions are true
• uncertain input parameters limit the interpretation of results
5
Why do we model?
• To inform decisions about resource allocation
• Therefore models should deliver:
– expected costs and health effects – for all options
– relating to appropriate population and sub-populations – based on full range of existing evidence
– quantification of decision uncertainty – valuation of future research
• In timely manner to support decisions
Schulpher M. ISPOR, Athens 2008
Modelling complementary to prospective approach
• intermediate to final outcome
• beyond trial duration
• beyond trial setting (costs and outcomes)
• compliance patients and physicians
• unobserved costs
• variability of costs, learning curves
• adaptation to new countries
6
Approximations are unaviodable in modelling
• ”Remember that all models are wrong: the practical question is how wrong do they have to be to not to be useful*”
• The search for ‘absolute accuracy’:
– adds complexity
– imposes costs (evidence gathering, computation time) – complicates communication
– increases potential modelling errors – need to justify in terms of better decisions
*Schulpher M. ISPOR, Athens 2008 based upon Box and Draper (1987) Empririlcal Model- Building and Response Surfaces p.424 Wiley
Application of models
• Decision analysis (clinical or economic)
– comparison of alternative treatment options/scenarios
• Prognostic models (epidemiology)
– probability estimates – pl. risk of disease/major events – intermediate endpoint hard endpoint
• Health policy estimates
– morbidity, mortality, health care spending for specific patients / diseases – prediction of major events to capacity or budget planning
7
Type of models
• Decision tree model
• Markov model
• Simulation model
– microsimulation (simulation of patients with individual characteristics) – discrete event simulation (time spent between events is calculated instead
of health status)
When to use decision tree models?
• Disease can be described with mutually exclusive patient routes.
• Transition of patients to different routes is based upon well-defined probabilities.
• Timing of patients’ transition in their routes is not important , timing of major clinical events within a route has no relevance (therefore decision tree models are applied for modelling exercises with short time horizon).
• Each patient route results in well-defined costs and outcomes.
Decision tree model
• Based on calculation of expected value
• Structure/nodes:
– decision nodes – probability nodes – end-points
8
• Each patient route (branch) represents future events with specific outcomes – health status
– costs
Steps of decision tree modelling
• Structure of decision tree
• Probabilities
• Outcomes related to end-points (e.g. QALY and cost)
• Expected value (backward calculation)
• Testing results
• Analysis of uncertainty
• Application of decision rule/decison
probability of no complication influenza probability of minor complication
probability of hospitalization yes
probability of death no influenza
Vaccination
probability of no complication influenza probability of minor complication
probability of hospitalization no
probability of death no influenza
9
Calculation of probabilities
• Primary sources (clinical trial, primary database analysis)
• Secondary sources (publication, aggregation of information from different sources, expert opinion)
• Point estimate
+ confidence interval
+ distribution (stochastic analysis)
• Caveat: sum of probabilities after probabilistic nodes = 1
Prevention of surgical infection
Cost of surgery: € 1000
No infection
• Length of stay: 1 ICU day + 5 day normal ward
• Cost of ICU days: €500
• Cost of normal ward: €100
Infection
• Probability: 15%
• Length of stay: 4 ICU days + 7 days normal ward
• Extra medication: €400
New generation antibiotic drug
• reduces risk of infection by 40%
• cost: €120
Would you use this drug in your hospital?
10
Vascular surgery Exercise
• The current task is the cost-effectiveness analysis of alternative treatment options of patients with a vascular disease1. Patients are 45-50 year-old males, for whom there are 3 different medical alternatives.
– no medical treatment
– medical management (drug therapy)
– vascular surgery: vascular implant + adjuvant drug therapy. Surgery does not always deliver the most optimal outcome, only 89% of patient gain significant
Old prevention New prevention
No infection Infection No infection Infection
Probability 85% 15% 91% 9%
ICU Normal ICU Normal ICU Normal ICU Normal
Length of stay 1 5 4 7 1 5 4 7
Cost of
hospitalization 500 € 500 € 2 000 € 700 € 500 € 500 € 2 000 € 700 €
Cost of drugs 400 € 120 € 520 €
Total costs 1 315 € 1 309 €
11 health benefit from the operation, health status of additional 10,5% remains unchanged compared to their health status before the surgery. Surgery is not risk-free, 0,5% of patients die during the surgery.
1 Information and data in the exercise is hypothetical, therefore cannot be cannot be used for decision-making purposes in real-life.
Answer the below questions:
• Which therapy is the cheapest and the most expensive for the sickness fund?
• Which therapy results in the most health gain for the patients?
• Which therapy is the more cost-effective compared to the no medical therapy alternative, the medical management or the vascular surgery?
Expected life years:
• no medical treatment: 5 years
• medical management: 9 years
• successful surgery: 15 years
• unsuccessful surgery (as medical mgmt): 9 years
Utility weights
Average utility weight for each life years until death.
• no medical therapy: 0.5
• medical management: 0.6
• successful surgery: 0.7
• unsuccessful surgery (as medical mgmt): 0.6
Cost
• no medical treatment: 0 €/year
• Medical management: 650 €/year
• adjuvant drug therapy to vascular implant: 200 €/year
• Surgery + vascular implant: 7450 €
• unsuccessful surgery (as medical mgmt): 650 €/year
• Discount rate (for the sake of simplicity) is 0%
12
• Would you recommend that the health insurance fund should reimburse the vascular surgery? If yes, why?
Vascular surgery
probability cost life
years utility
no treatment 100% 0 5 0.5
medical management 100% 5850 9 0.6 successful vascular
surgery 89% 10450 15 0.7
unsuccesful vascular
surgery 10.5% 13300 9 0.6
death 0.5% 7450
cost QALY
ICER =
∆cost /
∆QALY
no treatment 0 2.5
medical
management 5850 5.4 2017
vascular surgery 10734 9.9 1448
13
Markov model
• Different health states of particular disease, health states are mutually exclusives.
• Describes the possible transitions among health states.
• Transition among health states is based upon well-defined probabilities.
• Each health state has well defined cost and outcome (e.g quality of life).
• Consists of several monthly/yearly cycles.
• (No patient history, outcome in each cycle is independent from number of previous cycles).
Healthy
100
Healthy
70
Healthy
51
Diseased
0
Death
0
Diseased
20
Diseased
26
Death
10
Death
23
t+1
t+2
0.2 0.1
0.1
0.6 0.3 1
0.7
t
14
Markov model: prevention
Markov model
15
Markov model: curative therapy
Markov model: kidney transplantation
16
Markov states and transitions
Health states
• Mutually exclusives
• Clinically and economically important states
• Absorbing states (no return)
Length of Markov cycles
• Equal length of cycles
• Determining the length of cycles depends on
17 – the length of clinical events
– depth of available information – question to be answered
• How to calculate outcomes during transition among Markov cycles – according to outcomes of the new health state?
– according to outcomes of the old health state?
– half cycle correction (the longer cycles, the better to use half cycle correction)
Transition probabilities
• Markov chain: stable transition probabilities
• Markov process: transition probabilities can change over time/depend on the number of previous cycles
• Sum of transition probabilities after each health state = 1
• Primary sources (clinical trial, primary database analysis)
• Secondary sources (publication, aggregation of information from different sources, expert opinion)
• Point estimate
+ confidence interval
+ distribution (stochastic analysis)
18
Initial distribution of patients
• Incidence model: every patient in the cohort starts from the same health state in the first cycle
• Prevalence model: patient cohort starts from the different health states according to predefined distribution in the first cycle
Steps of Markov modelling
• Structure of Markov model – definition of health states – transition among health states
– length of Markov cycles (half cycle correction?)
• Probabilities (transition probabilities + initial distribution of patients)
• Outcomes in each health state (e.g. QALY and cost)
• Discounting
• Calculation of expected value
• Testing results
• Enalysis of uncertainty
• Application of decision rule/decision
19
Missing transition probabilities
• Look for primary of secondary data sources
• Develop algorithm from best available data sources (half-life, expected survival etc.)
• Calibrate the model to reflect the most important reference points (e.g. 1 and 5 year progression free survival)
Markov model
• cohort simulation
• fixed cycle length
• transition between states
• transition probabilities (with or without memory)
• costs and outcomes for each health state
• less input data needed
Microsimulation model
• patient level simulation
• fixed cycle length
• transition between states
• individual transition probabilities (with memory)
• costs and outcomes for each health state
• more input data needed
20 Markov model
• cohort simulation
• fixed cycle length
• transition between states
• transition probabilities (with or without memory)
• costs and outcomes for each health state
• less input data needed
Discrete event simulation
• patient level simulation
• focus on time to event
• variability among patients in number of different events
• costs and outcomes for each event
• huge and complex input data needed